News · 2026-08-08

WeatherNext called Melissa's Category 5 landfall five days out

Google DeepMind's WeatherNext model predicted that Hurricane Melissa would make landfall in Jamaica as a Category 5 storm five days in advance, with 80% confidence, rising to near-certainty three days out. On the harder question of where a storm will go, DeepMind reports that its five-day track forecasts landed on average about 140 kilometres closer to the truth than the European ensemble forecasters have relied on for decades -- a gap the company translates into roughly a day and a half of extra warning.

Key facts

• About 140 km closer on five-day cyclone track than the European ensemble, measured on the hurricane centre's 2023-2024 test years; that matches the ensemble's three-and-a-half-day skill.
• Hurricane Melissa: a Category 5 Jamaican landfall called five days ahead at 80% confidence.
• WeatherNext 2 forecasts now feed Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API, with raw data in Earth Engine, BigQuery, and Vertex AI.
• Primary sources: DeepMind's Melissa write-up, the WeatherNext 2 announcement, and NOAA's agreement with Google.

Rapid intensification is the failure mode that kills people. A storm that strengthens two categories in a day outruns evacuation orders that were issued against a weaker forecast. It is also the case conventional models handle worst, because it depends on small-scale processes inside the storm's core that a coarse simulation smooths away. Calling Melissa's Category 5 landfall five days out, with a confidence number attached, is the single most operationally meaningful result in the set -- more than any track statistic, because it is the one that changes what officials decide to do.

The mechanism is a genuine departure from how forecasting has worked since the 1950s. A conventional model discretises the atmosphere into a grid and numerically integrates the equations of fluid motion forward in time. WeatherNext does not solve those equations at all. It is a neural network trained on decades of historical weather that learned to map the current atmospheric state onto the next one. In the technical write-up, Skillful joint probabilistic weather forecasting from marginals, DeepMind describes a Functional Generative Network trained to score well on each weather variable individually, then recovering the relationships between variables through structured noise injected into the model itself plus ensembling across many runs.

That last part is the subtle bit, and it is why the confidence numbers exist at all. The model does not produce one forecast. It produces hundreds of plausible futures from a single starting point, and the "80% confidence" figure is simply the fraction of those futures in which Melissa arrives as a Category 5. This is the ensemble idea that has underpinned probabilistic forecasting for thirty years, done with a network fast enough to make hundreds of runs cheap. DeepMind says WeatherNext 2 surpasses its predecessor on 99.9% of variables and lead times while generating forecasts eight times faster, out to 15 days and, in an experimental mode, at hourly resolution.

The forecasts are no longer confined to a research page. Under a cooperative research and development agreement, NOAA says Google will supply near-real-time AI tropical cyclone forecasts to the National Hurricane Center so it can evaluate the models quickly and feed back improvements. Meanwhile the same underlying data ships inside consumer Google products and through developer surfaces including the WeatherNext model APIs.

Now the brake pedal, which comes from the forecasters rather than from critics. The National Hurricane Center's public Q&A on AI in hurricane forecasting says its official forecast remains the most skillful and consistent overall, that AI models are complementary rather than a replacement, and -- importantly -- that there were storms in the 2025 season where traditional models did better. The agency treated 2025 as a verification and trust-building period before wider integration. DeepMind's own Weather Lab carries the same disclaimer: it is a research surface, and live predictions there are not official warnings.

There is one more limit worth naming, which follows from the design rather than from anything DeepMind published. A model that learned the atmosphere from the past has only ever seen the past. Physics-based simulation has no such dependency: the equations do not care whether a situation is unprecedented. As the climate shifts the distribution of storms away from the historical record, a learned forecaster is extrapolating in exactly the regime where its training gives it least support -- and rare, record-breaking storms are precisely the ones a warning system exists for. Nobody has yet published a clean measurement of how much that costs.

Even so, the read for the next hurricane season is straightforward. The useful claim is not that AI beat physics. It is that on the hardest, most expensive failure in hurricane forecasting -- seeing explosive intensification early enough to act -- a learned model gave forecasters something they did not have before, and the agency responsible for the warnings is being careful about exactly how much weight to put on it. The Hacker News discussion of the WeatherNext 2 release, which drew 291 points and 131 comments, is a good place to watch working meteorologists argue the same point.

Primary source, verified: read the paper → (arXiv 2506.10772)

Ground Truth. (2026, August 8). WeatherNext called Melissa's Category 5 landfall five days out. Ground Truth. https://groundtruth.day/news/weathernext-called-melissas-category-5-landfall-five-days-out.html

@misc{groundtruth:weathernext-called-melissas-category-5-landfall-five-days-out,
  title  = {WeatherNext called Melissa's Category 5 landfall five days out},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/weathernext-called-melissas-category-5-landfall-five-days-out.html}
}

Topics: deepmind · weather · science · forecasting · google · ensembles